Blind Separation Of Multiple Speech Signals

Separation of speech mixtures, often referred to as the cocktail party problem, has been studied for two decades. In many source separation tasks, the separation method is limited by the assumption of at least as many sensors as sources. Further,many methods require that the number of signals within the recorded mixtures be known in advance. In many real-world applications these limitations are too restrictive.We have proposed a novel method for separating instantaneous and anechoic mixtures with an arbitrary number of speech signals of equal power with only two microphones.We have dealt with underdetermined mixtures by applying ICA to produce independent subsets. The subsets are used to estimate binary T-F masks, which are then applied to separate original mixtures. This iterative procedure continues until the independent subsets consist of only a single source. The segregated signals are further improved by merging masks from correlated subsets.Extensive MATLAB evaluation shows that mixtures of up to seven speech signals under anechoic conditions can be separated. The estimated binary masks are close to the ideal binary masks. The proposed framework has also been applied to speech mixtures recorded in a reverberant room.We find that instantaneous ICA applied iteratively in the time domain can be used to segregate convolutive mixtures. Through comparison In summary, our comparison with DUET suggests that the proposed method produces better results for instantaneous mixtures and comparable results for convolutive mixtures.Last we analysis the limitation of the paper and suggest two method to inprove.one is based on the local frequendy and another is the optimization of Parameters and threshold.Further search is on processing.